ctt-shell

Universal AI agent framework that makes 3B parameter models compose and execute multi-step plans like 12B models.

The thesis: structured context at inference time substitutes for parameter count — *Context-Time Training*.

[](https://nodejs.org/) [](https://www.typescriptlang.org/) [](https://opensource.org/licenses/MIT) []()

 "Create a post, assign a         ┌─────────────┐
  category, and upload media" ───→ │  CTT Memory  │ ← Knowledge, Skills, Memories
                                   └──────┬──────┘
                                          ▼
                                   ┌─────────────┐
                                   │  LLM (3B+)  │ ← Few-shot context + anti-patterns
                                   └──────┬──────┘
                                          ▼
                                   ┌─────────────┐
                                   │ Guard Rails  │ ← Normalize JSON, fix deps, retry
                                   └──────┬──────┘
                                          ▼
                                   ┌─────────────┐
                                   │   Execute    │ ← Domain adapter (WordPress, n8n, ...)
                                   └──────┬──────┘
                                          ▼
                                   ┌─────────────┐
                                   │    Learn     │ ← Save as Skill or Memory
                                   └─────────────┘

An agent receives a natural language goal, searches its CTT memory for relevant operations and patterns, asks any LLM to generate a plan, normalizes the output through battle-tested guard rails, executes it via a domain adapter, and learns from the